Search “Reasoning”

NVIDIA uses a 12,000-line Megakernel to optimize Kimi K3 inference, and vLLM says AI code is now faster than human review

Benjamin Chislett, an engineer at NVIDIA CentML, used a 12,000-line Megakernel written in CuTe DSL to optimize Moonshot AI's Kimi K3 inference. The vLLM blog reports over 840 tokens/s at batch 1 and a 1.5x to 2x increase in decoding throughput across multiple configurations, while SemiAnalysis estimates 423 tok/s/user on GB300 NVL72, a 1.68x improvement. The vLLM blog says agent-optimized inference code is now outpacing human review and maintenance speed, so the team proposed a fast-track model optimization RFC that lets reviewers judge designs from high-level abstractions rather than reviewing code line by line. The article also mentions that NVIDIA is developing optimizations for other models for vLLM, such as DeepSeek V4.1.

MaaS Has Arrived: Anthropic Turns an Operating Profit for the Quarter, China's Public Cloud MaaS Revenue Reaches 10.63 Billion Yuan in Half a Year

Anthropic's revenue in the second quarter of this year was 11.555 billion US dollars, achieving an operating profit of 974 million US dollars for the first time, with about 80% of revenue coming from pay-as-you-go billing. IDC data shows that in the first half of 2026, China's public cloud MaaS market revenue reached 10.63 billion yuan, up 889.4% year on year, and its share of public cloud PaaS revenue rose from 5.2% in the second half of last year to 21.0%. Zhipu's open platform and API revenue was 825 million yuan, accounting for 86.5% of total revenue; Volcano Engine ranked first with a 45.1% share. The article argues that model services can now generate gross profit, and the unit of competition has shifted to inference optimization and delivery capability.

What happened in the embodied intelligence industry 50 days after the release of GPT-6 Astra

TMTPost's Digital Intelligence Frontier reviewed the impact on the embodied intelligence industry in the roughly 50 days since OpenAI released GPT-6 Astra: Astra can directly control a robotic arm to complete tasks such as drawing a bridge without training on specialized robot data, and Nvidia's Jim Fan said it pushed the benchmark of his company's robotics framework to about 99%. However, an evaluation by SenseTime together with multiple institutions concluded that while it already has the ability to complete embodied tasks, there is still a clear gap before practical use: a single inference takes an average of 18.56 seconds, and the task of putting a pen into a pen holder took more than 4 hours and cost 55 US dollars. Three usage patterns have emerged in the industry: directly serving as a policy model, reviewing and correcting VLA actions, and building simulation environments and data annotation; Unitree's Wang Xingxing said that next year GPT-7 may directly break through the embodied intelligence problem.

Anthropic Races Toward a Trillion-Dollar Valuation While Zhipu Focuses on API Gross Margin: A Comparison of the Commercialization Paths of Chinese and U.S. Large-Model Companies

TMTPost's Large Model Home compared Anthropic's prospectus with Zhipu's interim report, pointing out that Chinese and U.S. large-model companies are taking two different commercialization paths. In the first half of the year, Anthropic reported revenue of $16.29 billion and a net loss of $80.7 billion, of which $72.4 billion came from fair value losses on convertible notes caused by rising valuations, while its operating loss narrowed to $693 million; Zhipu reported revenue of 954 million yuan and a loss of 2.072 billion yuan in the first half, with its API business gross margin turning positive for the first time, reaching 24.6%. The article also mentioned that Anthropic's computing power procurement commitments over the next decade amount to $518 billion, about 80% of which is non-cancelable, while Zhipu's per-token inference cost dropped by 80% compared with the beginning of the year.

Chinese Open Source Models Enter the Enterprise: Red Hat Uses GLM to Handle 70% of Inference Traffic

AI Business reported that, in the week ending October 25, Chinese models accounted for more than half of the token statistics on OpenRouter, whereas about a year ago only one of the nine models on the site's weekly usage list came from China. Red Hat is internally piloting self-hosted GLM, used by up to 1,400 engineers, with about 70% of inference traffic going to open source models, and it controls costs with a frontier model quota of $300 per person per month. DataCamp, due to latency and insufficient capacity from open source model providers, is shifting from Anthropic to OpenAI and has not yet moved its AI tutor to open source models.

Huang Renxun says doubling model speed can double AI lab revenue

Huang Renxun stated that, under the premise that demand exceeds capacity, doubling model speed (for example, through software or kernel optimization) can double AI lab revenue. Inference revenue equals the number of tokens served multiplied by the price per token; faster kernels can increase tokens per second under the same GPUs, power, and data center conditions. If all additional tokens can be sold at the current price, a 2x speedup can double revenue with zero additional capital expenditure.

Amazon CEO Jassy on $220 Billion in Capital Expenditure: Compute Is Still in Short Supply, Inference Costs Need to Fall by Another Order of Magnitude

Amazon CEO Andy Jassy said in an interview on the a16z podcast that the company's capital expenditure is expected to reach $220 billion in 2026, but compute capacity is already sold out for 2026, essentially fully booked for 2027, and a large share of 2028 quotas has been locked in. He described enterprise AI demand as dumbbell-shaped: a large number of production systems in the middle have yet to be deeply integrated because inference costs are too high, and costs need to fall by another order of magnitude before an explosion in adoption occurs. Jassy also noted that 99% of AWS's top 1,000 customers use Graviton, which offers about 40% better price-performance than x86, while Trainium has a 30% to 40% advantage over comparable hardware.

Amazon CEO Jassy: AI inference costs need to fall to about one-tenth of current levels

In an a16z interview, Amazon CEO Andy Jassy said AI inference costs are still an order of magnitude too high, and that for AI to be widely used in enterprise software, costs need to fall to about one-tenth of current levels. He said Amazon plans to spend $220 billion in capital expenditure this year, that its main compute supply for 2026 is essentially fully booked, and that most of 2027 has also been reserved; the in-house Trainium chip offers about 30% to 40% better price-performance than comparable processors, and has already secured more than $225 billion in future revenue commitments.

Nathan Lambert: DeepSeek processes 150T tokens per day on a single cluster, more than the combined total of several open-source inference platforms

Nathan Lambert says DeepSeek processes 150T tokens per day on a single cluster, more than the combined open-source model inference volume on OpenRouter, Fireworks, Together, and BaseTen. He also posted a reconstructed comparison chart of his own based on these companies' inference volume reports, and noted that this figure corresponds to only a single cluster; it remains unclear how many clusters DeepSeek actually has.

English translation is being prepared.

Leash Bio's Ian Quigley on DEL data, model cheating, and an agentic chemistry reasoning sandbox

In a Latent Space podcast interview, Leash Bio co-founder Ian Quigley discusses how the company tamed DNA-encoded libraries (DEL) and extracted clean, reproducible protein-small molecule interaction data from them. The interview mentions that in its Kaggle competition, none of the roughly 2,000 teams surpassed random performance on an unseen chemical library, and that a model can guess a chemist's identity from molecular structure, while Hermes, using only its own data, matched Boltz on public benchmarks. Ian also previewed that the team is turning years of accumulated data into tens of thousands of lab-in-the-loop simulations for agents to perform verifiable scientific reasoning; that work later developed into Minos, with a preprint, blog post, and open dataset released.

English translation is being prepared.

Inference spending to surpass training for the first time in 2026, rise of Token factories to change AI infrastructure investment logic

Gartner predicts that global inference spending will reach $23.3 billion in 2026, surpassing training spending of $19 billion for the first time and accounting for 55% of total AI IaaS spending; a CAICT report predicts that in 2026 inference computing power will account for 70.5% of total AI computing load. Multiple industry insiders point out that Hugging Face weight downloads and OpenRouter Token call volumes are often conflated as overseas expansion results, but Token call volume does not equal revenue; Token going global is essentially inference services going global, which is more directly related to actual revenue. Public cloud giants, independent MaaS inference service providers, and GPU vendors are all laying out Token factories, with compliance and global low-latency distribution being the main current bottlenecks.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

SemiAnalysis Analyzes Anthropic S-1: 90% Incremental Gross Margin and Half-Trillion Compute Commitment

The SemiAnalysis podcast analyzes Anthropic's S-1, pointing out its 90% incremental gross margin, 40% incremental EBIT, 85% inference gross margin, and 41% compute gross margin, and that it has been profitable for two consecutive quarters on a non-GAAP basis. On customer net revenue retention, customers who spent $1 last year now spend $9, with the Q1 figure at 500%; in addition, 250 customers pay over $10 million annually and 1,900 pay over $1 million. The episode also breaks down the $500 billion to $600 billion compute commitment, which at $100 billion per gigawatt translates to only about 5 gigawatts, and discusses the discrepancy in reporting that the top five customers account for 25% of revenue and that 48% goes through cloud vendor marketplaces.

AI agents need to be replaceable without losing approved business operations

The New Stack published an article discussing how to safely take over work after an AI agent crashes: before dispatching a refund, the application first writes the amount, authorization, idempotency key, and status to transactional storage, and the replacement agent inherits the original operation rather than re-reasoning. The article points out that durable execution frameworks such as LangGraph and Temporal can restore progress, but cannot determine whether the payer has already submitted the request; therefore, persistent records must be validated at the tool boundary, and the MCP services of Langfuse and ClickStack should be used to read-only investigate calls with unclear results.